fix: P0 hardening — fail-closed SECRET_KEY, locked CORS, role-safe registration (HARDENING.md P0.1/P0.2/P0.3)

P0.1 — fail-closed secrets:
- config.py: no default SECRET_KEY; refuses to boot when unset, a known
  placeholder, or <32 chars. Generate with: openssl rand -hex 32.
- docker-compose.yml: literal secrets removed; runtime env now comes from
  a git-ignored .env via env_file. .env.example added as template.
- .gitignore already covers .env (verified).

P0.2 — locked CORS:
- main.py: CORS_ORIGINS must be an explicit comma-separated allow-list.
  '*' or an empty value refuses to boot (was: silently ['*'] with
  allow_credentials=True).

P0.3 — role-safe registration:
- services/auth.py: client-supplied 'role' is IGNORED on POST
  /api/auth/register; self-registered users always get the
  least-privilege 'CS Rep' role. Unauthenticated callers can no longer
  mint Admin/Jerome, Admin/Wahab, or Director accounts.

Tests:
- conftest.py sets test SECRET_KEY/CORS_ORIGINS before app import.
- New tests/test_p0_hardening.py (8 tests): role-escalation blocked for
  Admin/Jerome and Admin/Wahab, duplicate-email 409, and subprocess
  boot-validation for placeholder/short/missing secret + wildcard CORS.
- Full suite: 44 passed.

Redeploy note (per research): seed_units/seed_categories are insert-only,
so the Aug-26 redeploy does NOT orphan historical tickets referencing
units 103E/103W/105E/105W or the legacy 34-category tree. Pending
Wahab: are 103E/103W/105E/105W real apartments dropped from the Excel
regeneration? Optional follow-up: floor-number backfill for already-
seeded units (mapping corrected floors; existing rows keep old values).

Checks per HARDENING.md acceptance:
- [x] starting without a real key fails loudly (subprocess-verified)
- [x] compose carries no literal secret; secrets come from .env
- [x] CORS_ORIGINS explicit allow-list, '*' rejected
- [x] unauthenticated register cannot mint Admin/* or Director

app/core/config.py

@@ -23,7 +23,7 @@ class Settings(BaseSettings):
     DATABASE_URL: str = "sqlite+aiosqlite:///./denya_onecare.db"
 
     # ── Auth ─────────────────────────────────────────────────────────
-    SECRET_KEY: str = "change-me-in-production-use-a-real-secret"
+    SECRET_KEY: str = ""
     ALGORITHM: str = "HS256"
     ACCESS_TOKEN_EXPIRE_MINUTES: int = 60  # Phase 1: raised 30 -> 60 for fewer re-logins
     REFRESH_TOKEN_EXPIRE_MINUTES: int = 60 * 24 * 7  # 7 days
@@ -42,3 +42,21 @@ class Settings(BaseSettings):
 
 
 settings = Settings()
+
+# ── Fail-closed secret validation (HARDENING.md P0.1) ─────────────────
+# Refuse to boot without a real SECRET_KEY. Devs must create a local .env
+# (see .env.example); production injects it via docker-compose env_file.
+_KNOWN_PLACEHOLDER_SECRETS = {
+    "",
+    "change-me-in-production",
+    "change-me-in-production-use-a-real-secret",
+    "changeme",
+    "secret",
+}
+
+if settings.SECRET_KEY in _KNOWN_PLACEHOLDER_SECRETS or len(settings.SECRET_KEY) < 32:
+    raise RuntimeError(
+        "SECRET_KEY is missing, a known placeholder, or shorter than 32 chars. "
+        "Generate one with: openssl rand -hex 32 — and set it in .env "
+        "(dev) or the runtime environment (prod). Refusing to start."
+    )